Autonomous Modulation Classification using Single Inception Module based Convolutional Neural Network

Modulation recognition plays a crucial role in noncooperative communication in which receivers have no prior information regarding transmitter modulation scheme. This paper presents a novel convolutional neural network with single Stem and Inception module for autonomous modulation classification from raw I/Q received channels. This addresses the computationally intensive problem of conversion of I/Q channels to constellation image processing. The proposed system is capable of classifying 11 standard modulation schemes with both 2D and 3D input array configurations for varying SNR conditions. The performance of proposed system has been evaluated for a realtime communication system simulated with Rician fading channel and AWGN noise model, providing realistic distortion effects. The proposed CNN design achieves an average accuracy of 90% at 10 dB and 99% at 20 dB SNR with reduced network learnables and lower training and testing time, which makes it computationally efficient. These attributes make Inception module based CNN a viable solution for modulation classification in practical low cost and portable yet reliable communication systems.